Daily 9 for November 21, 2026 covers nine developments across AI, fintech, crypto, software and tech, announced between November 18 and November 20. Anthropic closed its initial public offering, OpenAI described how it defends against model distillation, Google launched Tool Retrieval in the Gemini API, and the SEC proposed taking direct control of the Consolidated Audit Trail.

Each item links to the company, project or agency that published it. Weekend news is thinner, so this edition reaches back to November 18 and states every date.

1. Anthropic closes its IPO at $83 a share

Anthropic announced the closing of its initial public offering of 1,127,000,000 Class A shares at $83.00, including the full exercise of the underwriters’ option. Of those, 1,103,351,354 shares were sold by the company and 23,648,646 by selling stockholders, for gross proceeds of about $93.5 billion, of which about $91.6 billion went to Anthropic. A concurrent private placement of 313,253,007 Class N shares to AMD, Broadcom, Google, Micron, Microsoft, NVIDIA and TSMC Partners brought in about $26.0 billion. Shares began trading on Nasdaq as “ANT” on November 19. Anthropic, November 20.

2. OpenAI details its defenses against adversarial distillation

OpenAI published how it detects and disrupts adversarial distillation, in which one model is trained on another’s outputs without authorization. It says the average attack success rate for recovering protected reasoning fell from 54.0% against GPT-6 Astra to 0.2% against GPT-6.2 Astra in an internal red-teaming evaluation. The safeguards combine classifiers, API limits that tie protected reasoning to the conversation that produced it, and investigations of account networks. OpenAI also says it shares indicators of compromise with other providers. Background: what model distillation is. OpenAI, November 20.

3. Gemini API adds Tool Retrieval in preview

Google launched Tool Retrieval in the Interactions API of the Gemini API. A developer adds a tool_search tool and sets defer_loading to true on tools to hold back, so the model sees each tool’s name and description but loads its full parameter schema only when needed. Google says this is meant to cut input tokens and latency for agents with large tool libraries. It works with Gemini 3.6 and later, including Gemini 4 models, and with remote MCP servers. There is no separate charge beyond standard token rates. Background: AI agents vs. chatbots. Google, November 19.

4. Microsoft Edge ships Aion Instruct as its built-in small model

Microsoft made Aion Instruct generally available as an open-weights model on Hugging Face and as Edge’s built-in small language model, replacing Phi-4-mini. Microsoft describes it as a 2.3-billion-parameter model comparable in quality for common web tasks. It also opened an origin trial of the Prompt API so live sites can call the on-device model, and put an experimental Decisions API in developer preview in Edge Canary and Dev. Edge downloads the model only when a site first needs it, and users can turn on-device AI off. Background: small language model vs. LLM. Microsoft Edge Blog, November 19.

5. Cloudflare releases Clairic extraction models as open weights

Cloudflare released an updated Clairic, a 27-billion-parameter extraction model, and Clairic-flash, a 9-billion-parameter version, with open weights on Hugging Face and hosting on Workers AI. The updated models read PDFs and up to four images or pages per request. Until December 1 both cost $0.10 per million input tokens; after that Clairic goes to $0.24 and Clairic-flash stays at $0.10, with no charge for output tokens. Cloudflare’s accuracy figures come from its own internal evaluations, which it calls directional. Cloudflare, November 18.

6. Google Quantum AI tests contextuality at scale

Google Quantum AI published a paper on the cover of Science showing that contextuality, a quantum property researchers believe powers certain algorithms, survives at scale on its processor. In a coordination game its qubits won 98.3% of the time, against an 88.9% classical limit, and beat constrained classical strategies in tasks spanning up to 105 qubits. Google says the tests also work as a sensitive way to catch subtle hardware errors. The result is a lab demonstration, not a practical application. Background: what a qubit is. Google, November 19; Science paper.

7. SEC charges eight entities over fake crypto trading platforms

The SEC charged eight Colorado-based entities, which it says were likely run by people overseas, with using fake crypto trading platforms to take more than $37 million from U.S. investors. According to the SEC, social media ads led victims into chat groups where people posing as an investment adviser gave stock advice, then pushed them to send money to a supposed crypto platform. The scheme ran as four adviser-platform pairs from August 2024 through May 2026. The SEC filed in U.S. District Court in Colorado and seeks injunctions, disgorgement and penalties. These are allegations. SEC, November 19.

8. SEC proposes to restructure the Consolidated Audit Trail

The SEC proposed rescinding Rule 613 of Regulation NMS, under which exchanges and national securities associations run the Consolidated Audit Trail, and adding Rule 17a-26. Exchanges, associations and broker-dealers would report substantially the same information to the Commission or its designee, putting the system under the SEC’s direct control. The agency says it intends to use funds appropriated by Congress to cover costs. It says earlier relief cut projected annual operating costs by over $100 million and ended reporting of personally identifiable information. Comments are open for 60 days after Federal Register publication. This is a proposal, not a final rule. SEC, November 20.

9. Bitcoin Core 30.5 and 31.3 fix a node-stall bug

Bitcoin Core released versions 30.5 and 31.3, which fix CVE-2026-117348, rated Medium severity. A peer could stall a node for over a dozen seconds per message by repeatedly announcing an orphan transaction with many missing parents, and during that time the node processes nothing from other peers, including blocks. Versions 29.0 to 29.5, 30.0 to 30.4 and 31.0 to 31.2 are affected; the 29.x series has reached end of life and gets no fix. The fix was already in 32.1. The project says an automated LLM scan found the flaw. Background: what a blockchain is. Bitcoin Core, November 20.